Document version: 2026-09-14
This document explains the optional cloud AI processing that you consent to in Nemosine. It applies only when you select Luna (Cloud), Grok (Cloud), or GLM 5.3 Flash (NEAR AI) and send a request. On-device AI is separate and sends nothing to Nemosine or a model provider.
1. What is sent
For each request, your device sends only the material assembled for that request. This may include:
- your prompt and relevant chat history
- note content or images that you chose to include
- available tool descriptions, tool calls, and tool results used to answer the request
- the selected model and generation settings
Free-form content may identify you or other people. Only send personal data that you are entitled to process and that you want the selected provider to handle.
2. Providers
The provider depends on the model you select:
- Luna (Cloud) uses OpenAI.
- Grok (Cloud) uses xAI (SpaceXAI LLC).
- GLM 5.3 Flash (NEAR AI) uses NEAR AI (Jasnah, Inc.).
The GLM model is identified as z-ai/glm-5.3-flash because its open weights originate with Z.ai. NEAR AI runs that model on its own hardware. Z.ai does not receive your request and is not a recipient of your content.
A request is sent only to the provider of the selected model. Nemosine uses a separate organisational API credential for each provider. We do not put your Nemosine user ID, email address, account details, or client IP address in the provider request.
This is pseudonymous processing, not guaranteed anonymity. The content itself may identify you.
3. Cache routing
Providers can temporarily reuse an unchanged prompt prefix to reduce processing time and cost.
For Grok, Nemosine may send an opaque per-conversation cache-routing value. Our server derives it with a keyed one-way function from your account and local conversation identifiers. xAI does not receive either underlying identifier and cannot use this value to link separate Nemosine conversations. It can recognise requests that carry the same derived value as belonging to the same conversation.
Nemosine currently sends no comparable explicit conversation cache key to OpenAI or NEAR AI. Both may still apply their own automatic prompt caching, and NEAR AI bills a reused prefix at a lower cached-input rate.
Reasoning state
Grok generates internal reasoning and does not offer an off setting. Nemosine requests its lowest available reasoning effort. We use xAI's stateless Responses API with server-side storage disabled and ask xAI to return that reasoning as an opaque, xAI-encrypted item. Nemosine cannot read or display the encrypted reasoning.
Your device stores the encrypted item, the exact request context that preceded it, and related tool state with the completed assistant message. A later request to the same Grok model can send those items back so Grok retains its earlier reasoning context and the unchanged prefix can qualify for prompt caching. This can reduce repeated work and cost, but it does not guarantee a cache hit or prevent Grok from reasoning about a new question. New reasoning tokens are still charged as output tokens.
If cloud sync is enabled, this state is included inside the same end-to-end encrypted sync row as the chat message. Otherwise it remains in local app data. Deleting the chat deletes Nemosine's local copy through the normal chat-deletion and sync process. Switching models, clearing history, or losing the state can remove the reuse benefit.
GLM produces internal reasoning text. When an answer uses tools, Nemosine holds that text in memory for the duration of that single answer and sends it back to NEAR AI between tool steps so the model keeps its own context. It is discarded when the answer completes. Unlike Grok state, it is never written to your chat history, never stored on your device, and never synced. Nemosine does not display it. New reasoning tokens are charged as output tokens.
Luna is requested with reasoning disabled because its current Chat Completions tool interface requires that setting. Nemosine does not receive reusable Luna reasoning state.
4. Provider retention and training
Nemosine does not opt in to provider model training with your cloud AI content.
- xAI: Nemosine enables Grok only through an xAI team with Zero Data Retention (ZDR). Under ZDR, request inputs and model outputs are not persisted to disk by xAI. Nemosine also verifies xAI's ZDR response header. Temporary in-memory processing and prompt caching may still occur while serving requests, and xAI may process operational metadata under its terms.
- NEAR AI (contracting entity: Jasnah, Inc.): GLM 5.3 Flash is an open-weights model that NEAR AI runs on its own GPUs inside a hardware Trusted Execution Environment (TEE). NEAR AI states that prompts and outputs processed inside that enclave cannot be read by the host operating system, the GPU operator, or NEAR AI itself. NEAR AI's Terms of Service state that it will not use customer data to train any generative AI model. NEAR AI processes personal data contained in that content under its Data Processing Agreement. A TEE reduces the risk of access to data while it is in use; it does not eliminate it. Hardware side-channel and firmware vulnerabilities are possible, and NEAR AI does not represent that TEE processing satisfies any particular regulatory requirement.
- OpenAI (contracting entity: OpenAI Ireland Limited): Nemosine requests that API responses are not stored as application state. OpenAI may nevertheless retain customer content in abuse-monitoring logs for up to 30 days unless Nemosine's OpenAI organisation is separately approved and configured for ZDR. OpenAI may also hold encrypted prompt-cache state temporarily on its GPU infrastructure.
NEAR AI's enclave protection applies to NEAR AI's own processing. It does not extend backwards to Nemosine. Your request reaches NEAR AI through Nemosine's relay, which necessarily handles the request in readable form in order to forward it. Selecting GLM does not give you encryption that runs unbroken from your device into NEAR AI's enclave, and Nemosine can see GLM content exactly as it can for the other models. What Nemosine does with it is described in section 5.
Response verification
For GLM, Nemosine asks NEAR AI for a cryptographic signature covering the exact bytes of the request it sent and the reply it received, and shows a checkmark next to the answer when both checksums and that signature verify.
A green checkmark means the request and reply were not altered in transit between Nemosine's relay and NEAR AI's signing key. It is not an independent verification of NEAR AI's hardware. Nemosine does not validate manufacturer attestation quotes, the software measured inside the enclave, the binding between the TLS session and the enclave, or the model weights. A checkmark that is grey, or that reports a failure, means only that a proof was unavailable or did not verify; it is not evidence that anything was tampered with.
The indicator reflects the most recent GLM answer in that chat during the current app session. It is not stored with your chat, not synced, and not shared between accounts. Retrieving a proof sends no additional content and costs no credits. Verification is independent of billing: an unavailable or failed proof does not make processing that already happened free.
Nemosine does not use provider-hosted conversation history, files, collections, or batch processing for this feature. Grok conversation state is client-managed as described above rather than stored by xAI. Nemosine does not use NEAR AI's stored-conversation or agent features.
5. What Nemosine stores
Our server function relays each request and does not store or log the prompt, included note content, image, tool result, or model reply. We store a usage and billing record containing token counts, selected model, timing, status, and calculated cost. Those records are retained for 24 months and then aggregated into statistics that do not identify you.
Your copy of the chat remains in Nemosine's local app data and follows the sync choices you make separately. For Grok, that local copy also contains the exact request context used for replay and xAI-encrypted reasoning/tool state. For GLM, the local copy contains the answer only; the reasoning text described in section 3 is not part of it. Synced copies are end-to-end encrypted before upload.
6. Consent and withdrawal
Cloud AI content processing is based on your explicit consent under GDPR Article 6(1)(a). Consent is optional. Refusing or withdrawing it does not affect on-device AI or your locally stored notes.
You can withdraw consent in Nemosine's settings. Withdrawal prevents future cloud AI requests but does not undo processing that already occurred or remove billing records that Nemosine must retain for the stated period.
See the Privacy Policy, GDPR information, and Terms of Service for the rest of Nemosine's data handling and your rights.